COMPMID-2805: Add QASYMM8_SIGNED support in NEGEMMLowpOutputStage
Add support from requantizing down from S32 to Int8 with fixed point
requantization. This involves the following:
- Compute fixed point multiplication between each entry of input by
result_fixedpoint_multiplier
- Add bias to final result if bias tensor is not a nullptr
- Round to nearest division by a power-of-two using result_shift
- Add offset to each result
- Clamp the value between the specified min and max bounds
- Cast to int8 data type
Change-Id: I641b3fac0833c568d8565ccb859bbc561a24c17d
Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-on: https://review.mlplatform.org/c/2340
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
diff --git a/arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h b/arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h
index 5ece753..1a65f3b 100644
--- a/arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h
+++ b/arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h
@@ -147,6 +147,65 @@
*/
static Status validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min = 0, int max = 0);
};
+/** Basic function to execute NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint on NEON.
+ *
+ * NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint depends on 3 parameters:
+ *
+ * result_fixedpoint_multiplier, result_shift, result_offset_after_shift
+ *
+ * The final result is:
+ *
+ * (FixedPointMul(input[i][k], result_fixedpoint_multiplier) >> result_shift) + result_offset_after_shift
+ *
+ * where FixedPointMul(x, y) is the nearest integer to the following
+ * mathematical expression, evaluated without overflow or intermediate rounding:
+ *
+ * (x * y) / 2^31
+ *
+ * For more information: https://github.com/google/gemmlowp/blob/master/public/output_stages.h#L68
+ *
+ * In case the bias tensor is provided, the final result is:
+ *
+ * ((FixedPointMul(input[i][k] + bias[k], result_fixedpoint_multiplier)) >> result_shift) + result_offset_after_shift
+ *
+ * This function calls the following NEON kernels:
+ *
+ * -# @ref NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel
+ *
+ * @note The function accepts also 2 optional input arguments (min and max) which can be used to implement "rectified linear unit" activation functions
+ * after the result is shifted right by result_shift
+*/
+class NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint : public INESimpleFunctionNoBorder
+{
+public:
+ /** Initialise the kernel's inputs, output
+ *
+ * @param[in] input Input tensor. Data type supported: S32
+ * @param[in] bias Biases tensor. Only shared biases supported and it can be a nullptr if the biases addition is not required.
+ * Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p input.
+ * @param[out] output Output tensor. Data type supported: Data type supported: QASYMM8_SIGNED
+ * @param[in] result_fixedpoint_multiplier Fixed point value to be multiplied to each element of the input matrix when once the result_offset has been add
+ * @param[in] result_shift Number of bits to shift right the result after the fixed point multiplication
+ * @param[in] result_offset_after_shift Offset to be applied to result before converting it back to QASYMM8_SIGNED
+ * @param[in] min (Optional) Min value used to saturate down the output result before converting back to QASYMM8_SIGNED
+ * @param[in] max (Optional) Max value used to saturate up the output result before converting back to QASYMM8_SIGNED,
+ * Along with @p min, this value can be used to implement "rectified linear unit" activation functions
+ */
+ void configure(const ITensor *input, const ITensor *bias, ITensor *output, int result_fixedpoint_multiplier, int result_shift, int result_offset_after_shift, int min = 0, int max = 0);
+ /** Static function to check if given info will lead to a valid configuration of @ref NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint
+ *
+ * @param[in] input Input tensor. It is the output of @ref NEGEMMLowpMatrixMultiplyCore function. Data type supported: S32
+ * @param[in] bias Biases tensor. Only shared biases supported and it can be a nullptr if the addition of biases is not required.
+ * Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p input.
+ * @param[in] output Output tensor. Data type supported: Data type supported: QASYMM8_SIGNED
+ * @param[in] min (Optional) Min value used to saturate down the output result before converting back to QASYMM8_SIGNED
+ * @param[in] max (Optional) Max value used to saturate up the output result before converting back to QASYMM8_SIGNED,
+ * Along with @p min, this value can be used to implement "rectified linear unit" activation functions
+ *
+ * @return a status
+ */
+ static Status validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min = 0, int max = 0);
+};
/** Basic function to execute NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint on NEON.
*
* NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint depends on 2 parameters: